1/*
2 * Copyright 2015 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23
24/**
25 * DOC: Overview
26 *
27 * The GPU scheduler provides entities which allow userspace to push jobs
28 * into software queues which are then scheduled on a hardware run queue.
29 * The software queues have a priority among them. The scheduler selects the entities
30 * from the run queue using a FIFO. The scheduler provides dependency handling
31 * features among jobs. The driver is supposed to provide callback functions for
32 * backend operations to the scheduler like submitting a job to hardware run queue,
33 * returning the dependencies of a job etc.
34 *
35 * The organisation of the scheduler is the following:
36 *
37 * 1. Each hw run queue has one scheduler
38 * 2. Each scheduler has multiple run queues with different priorities
39 * (e.g., HIGH_HW,HIGH_SW, KERNEL, NORMAL)
40 * 3. Each scheduler run queue has a queue of entities to schedule
41 * 4. Entities themselves maintain a queue of jobs that will be scheduled on
42 * the hardware.
43 *
44 * The jobs in an entity are always scheduled in the order in which they were pushed.
45 *
46 * Note that once a job was taken from the entities queue and pushed to the
47 * hardware, i.e. the pending queue, the entity must not be referenced anymore
48 * through the jobs entity pointer.
49 */
50
51/**
52 * DOC: Flow Control
53 *
54 * The DRM GPU scheduler provides a flow control mechanism to regulate the rate
55 * in which the jobs fetched from scheduler entities are executed.
56 *
57 * In this context the &drm_gpu_scheduler keeps track of a driver specified
58 * credit limit representing the capacity of this scheduler and a credit count;
59 * every &drm_sched_job carries a driver specified number of credits.
60 *
61 * Once a job is executed (but not yet finished), the job's credits contribute
62 * to the scheduler's credit count until the job is finished. If by executing
63 * one more job the scheduler's credit count would exceed the scheduler's
64 * credit limit, the job won't be executed. Instead, the scheduler will wait
65 * until the credit count has decreased enough to not overflow its credit limit.
66 * This implies waiting for previously executed jobs.
67 */
68
69#include <linux/export.h>
70#include <linux/wait.h>
71#include <linux/sched.h>
72#include <linux/completion.h>
73#include <linux/dma-resv.h>
74#include <uapi/linux/sched/types.h>
75
76#include <drm/drm_print.h>
77#include <drm/drm_gem.h>
78#include <drm/drm_syncobj.h>
79#include <drm/gpu_scheduler.h>
80#include <drm/spsc_queue.h>
81
82#include "sched_internal.h"
83
84#define CREATE_TRACE_POINTS
85#include "gpu_scheduler_trace.h"
86
87int drm_sched_policy = DRM_SCHED_POLICY_FIFO;
88
89/**
90 * DOC: sched_policy (int)
91 * Used to override default entities scheduling policy in a run queue.
92 */
93MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default).");
94module_param_named(sched_policy, drm_sched_policy, int, 0444);
95
96static u32 drm_sched_available_credits(struct drm_gpu_scheduler *sched)
97{
98 u32 credits;
99
100 WARN_ON(check_sub_overflow(sched->credit_limit,
101 atomic_read(&sched->credit_count),
102 &credits));
103
104 return credits;
105}
106
107/**
108 * drm_sched_can_queue -- Can we queue more to the hardware?
109 * @sched: scheduler instance
110 * @entity: the scheduler entity
111 *
112 * Return true if we can push at least one more job from @entity, false
113 * otherwise.
114 */
115static bool drm_sched_can_queue(struct drm_gpu_scheduler *sched,
116 struct drm_sched_entity *entity)
117{
118 struct drm_sched_job *s_job;
119
120 s_job = drm_sched_entity_queue_peek(entity);
121 if (!s_job)
122 return false;
123
124 /* If a job exceeds the credit limit, truncate it to the credit limit
125 * itself to guarantee forward progress.
126 */
127 if (s_job->credits > sched->credit_limit) {
128 dev_WARN(sched->dev,
129 "Jobs may not exceed the credit limit, truncate.\n");
130 s_job->credits = sched->credit_limit;
131 }
132
133 return drm_sched_available_credits(sched) >= s_job->credits;
134}
135
136static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
137 const struct rb_node *b)
138{
139 struct drm_sched_entity *ent_a = rb_entry((a), struct drm_sched_entity, rb_tree_node);
140 struct drm_sched_entity *ent_b = rb_entry((b), struct drm_sched_entity, rb_tree_node);
141
142 return ktime_before(cmp1: ent_a->oldest_job_waiting, cmp2: ent_b->oldest_job_waiting);
143}
144
145static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity *entity,
146 struct drm_sched_rq *rq)
147{
148 if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
149 rb_erase_cached(node: &entity->rb_tree_node, root: &rq->rb_tree_root);
150 RB_CLEAR_NODE(&entity->rb_tree_node);
151 }
152}
153
154void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity,
155 struct drm_sched_rq *rq,
156 ktime_t ts)
157{
158 /*
159 * Both locks need to be grabbed, one to protect from entity->rq change
160 * for entity from within concurrent drm_sched_entity_select_rq and the
161 * other to update the rb tree structure.
162 */
163 lockdep_assert_held(&entity->lock);
164 lockdep_assert_held(&rq->lock);
165
166 drm_sched_rq_remove_fifo_locked(entity, rq);
167
168 entity->oldest_job_waiting = ts;
169
170 rb_add_cached(node: &entity->rb_tree_node, tree: &rq->rb_tree_root,
171 less: drm_sched_entity_compare_before);
172}
173
174/**
175 * drm_sched_rq_init - initialize a given run queue struct
176 *
177 * @sched: scheduler instance to associate with this run queue
178 * @rq: scheduler run queue
179 *
180 * Initializes a scheduler runqueue.
181 */
182static void drm_sched_rq_init(struct drm_gpu_scheduler *sched,
183 struct drm_sched_rq *rq)
184{
185 spin_lock_init(&rq->lock);
186 INIT_LIST_HEAD(list: &rq->entities);
187 rq->rb_tree_root = RB_ROOT_CACHED;
188 rq->current_entity = NULL;
189 rq->sched = sched;
190}
191
192/**
193 * drm_sched_rq_add_entity - add an entity
194 *
195 * @rq: scheduler run queue
196 * @entity: scheduler entity
197 *
198 * Adds a scheduler entity to the run queue.
199 */
200void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
201 struct drm_sched_entity *entity)
202{
203 lockdep_assert_held(&entity->lock);
204 lockdep_assert_held(&rq->lock);
205
206 if (!list_empty(head: &entity->list))
207 return;
208
209 atomic_inc(v: rq->sched->score);
210 list_add_tail(new: &entity->list, head: &rq->entities);
211}
212
213/**
214 * drm_sched_rq_remove_entity - remove an entity
215 *
216 * @rq: scheduler run queue
217 * @entity: scheduler entity
218 *
219 * Removes a scheduler entity from the run queue.
220 */
221void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
222 struct drm_sched_entity *entity)
223{
224 lockdep_assert_held(&entity->lock);
225
226 if (list_empty(head: &entity->list))
227 return;
228
229 spin_lock(lock: &rq->lock);
230
231 atomic_dec(v: rq->sched->score);
232 list_del_init(entry: &entity->list);
233
234 if (rq->current_entity == entity)
235 rq->current_entity = NULL;
236
237 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
238 drm_sched_rq_remove_fifo_locked(entity, rq);
239
240 spin_unlock(lock: &rq->lock);
241}
242
243/**
244 * drm_sched_rq_select_entity_rr - Select an entity which could provide a job to run
245 *
246 * @sched: the gpu scheduler
247 * @rq: scheduler run queue to check.
248 *
249 * Try to find the next ready entity.
250 *
251 * Return an entity if one is found; return an error-pointer (!NULL) if an
252 * entity was ready, but the scheduler had insufficient credits to accommodate
253 * its job; return NULL, if no ready entity was found.
254 */
255static struct drm_sched_entity *
256drm_sched_rq_select_entity_rr(struct drm_gpu_scheduler *sched,
257 struct drm_sched_rq *rq)
258{
259 struct drm_sched_entity *entity;
260
261 spin_lock(lock: &rq->lock);
262
263 entity = rq->current_entity;
264 if (entity) {
265 list_for_each_entry_continue(entity, &rq->entities, list) {
266 if (drm_sched_entity_is_ready(entity))
267 goto found;
268 }
269 }
270
271 list_for_each_entry(entity, &rq->entities, list) {
272 if (drm_sched_entity_is_ready(entity))
273 goto found;
274
275 if (entity == rq->current_entity)
276 break;
277 }
278
279 spin_unlock(lock: &rq->lock);
280
281 return NULL;
282
283found:
284 if (!drm_sched_can_queue(sched, entity)) {
285 /*
286 * If scheduler cannot take more jobs signal the caller to not
287 * consider lower priority queues.
288 */
289 entity = ERR_PTR(error: -ENOSPC);
290 } else {
291 rq->current_entity = entity;
292 reinit_completion(x: &entity->entity_idle);
293 }
294
295 spin_unlock(lock: &rq->lock);
296
297 return entity;
298}
299
300/**
301 * drm_sched_rq_select_entity_fifo - Select an entity which provides a job to run
302 *
303 * @sched: the gpu scheduler
304 * @rq: scheduler run queue to check.
305 *
306 * Find oldest waiting ready entity.
307 *
308 * Return an entity if one is found; return an error-pointer (!NULL) if an
309 * entity was ready, but the scheduler had insufficient credits to accommodate
310 * its job; return NULL, if no ready entity was found.
311 */
312static struct drm_sched_entity *
313drm_sched_rq_select_entity_fifo(struct drm_gpu_scheduler *sched,
314 struct drm_sched_rq *rq)
315{
316 struct rb_node *rb;
317
318 spin_lock(lock: &rq->lock);
319 for (rb = rb_first_cached(&rq->rb_tree_root); rb; rb = rb_next(rb)) {
320 struct drm_sched_entity *entity;
321
322 entity = rb_entry(rb, struct drm_sched_entity, rb_tree_node);
323 if (drm_sched_entity_is_ready(entity)) {
324 /* If we can't queue yet, preserve the current entity in
325 * terms of fairness.
326 */
327 if (!drm_sched_can_queue(sched, entity)) {
328 spin_unlock(lock: &rq->lock);
329 return ERR_PTR(error: -ENOSPC);
330 }
331
332 reinit_completion(x: &entity->entity_idle);
333 break;
334 }
335 }
336 spin_unlock(lock: &rq->lock);
337
338 return rb ? rb_entry(rb, struct drm_sched_entity, rb_tree_node) : NULL;
339}
340
341/**
342 * drm_sched_run_job_queue - enqueue run-job work
343 * @sched: scheduler instance
344 */
345static void drm_sched_run_job_queue(struct drm_gpu_scheduler *sched)
346{
347 if (!READ_ONCE(sched->pause_submit))
348 queue_work(wq: sched->submit_wq, work: &sched->work_run_job);
349}
350
351/**
352 * drm_sched_run_free_queue - enqueue free-job work
353 * @sched: scheduler instance
354 */
355static void drm_sched_run_free_queue(struct drm_gpu_scheduler *sched)
356{
357 if (!READ_ONCE(sched->pause_submit))
358 queue_work(wq: sched->submit_wq, work: &sched->work_free_job);
359}
360
361/**
362 * drm_sched_job_done - complete a job
363 * @s_job: pointer to the job which is done
364 *
365 * Finish the job's fence and resubmit the work items.
366 */
367static void drm_sched_job_done(struct drm_sched_job *s_job, int result)
368{
369 struct drm_sched_fence *s_fence = s_job->s_fence;
370 struct drm_gpu_scheduler *sched = s_fence->sched;
371
372 atomic_sub(i: s_job->credits, v: &sched->credit_count);
373 atomic_dec(v: sched->score);
374
375 trace_drm_sched_job_done(fence: s_fence);
376
377 dma_fence_get(fence: &s_fence->finished);
378 drm_sched_fence_finished(fence: s_fence, result);
379 dma_fence_put(fence: &s_fence->finished);
380 drm_sched_run_free_queue(sched);
381}
382
383/**
384 * drm_sched_job_done_cb - the callback for a done job
385 * @f: fence
386 * @cb: fence callbacks
387 */
388static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
389{
390 struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
391
392 drm_sched_job_done(s_job, result: f->error);
393}
394
395/**
396 * drm_sched_start_timeout - start timeout for reset worker
397 *
398 * @sched: scheduler instance to start the worker for
399 *
400 * Start the timeout for the given scheduler.
401 */
402static void drm_sched_start_timeout(struct drm_gpu_scheduler *sched)
403{
404 lockdep_assert_held(&sched->job_list_lock);
405
406 if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
407 !list_empty(head: &sched->pending_list))
408 mod_delayed_work(wq: sched->timeout_wq, dwork: &sched->work_tdr, delay: sched->timeout);
409}
410
411static void drm_sched_start_timeout_unlocked(struct drm_gpu_scheduler *sched)
412{
413 spin_lock(lock: &sched->job_list_lock);
414 drm_sched_start_timeout(sched);
415 spin_unlock(lock: &sched->job_list_lock);
416}
417
418/**
419 * drm_sched_tdr_queue_imm: - immediately start job timeout handler
420 *
421 * @sched: scheduler for which the timeout handling should be started.
422 *
423 * Start timeout handling immediately for the named scheduler.
424 */
425void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched)
426{
427 spin_lock(lock: &sched->job_list_lock);
428 sched->timeout = 0;
429 drm_sched_start_timeout(sched);
430 spin_unlock(lock: &sched->job_list_lock);
431}
432EXPORT_SYMBOL(drm_sched_tdr_queue_imm);
433
434/**
435 * drm_sched_fault - immediately start timeout handler
436 *
437 * @sched: scheduler where the timeout handling should be started.
438 *
439 * Start timeout handling immediately when the driver detects a hardware fault.
440 */
441void drm_sched_fault(struct drm_gpu_scheduler *sched)
442{
443 if (sched->timeout_wq)
444 mod_delayed_work(wq: sched->timeout_wq, dwork: &sched->work_tdr, delay: 0);
445}
446EXPORT_SYMBOL(drm_sched_fault);
447
448/**
449 * drm_sched_suspend_timeout - Suspend scheduler job timeout
450 *
451 * @sched: scheduler instance for which to suspend the timeout
452 *
453 * Suspend the delayed work timeout for the scheduler. This is done by
454 * modifying the delayed work timeout to an arbitrary large value,
455 * MAX_SCHEDULE_TIMEOUT in this case.
456 *
457 * Returns the timeout remaining
458 *
459 */
460unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched)
461{
462 unsigned long sched_timeout, now = jiffies;
463
464 sched_timeout = sched->work_tdr.timer.expires;
465
466 /*
467 * Modify the timeout to an arbitrarily large value. This also prevents
468 * the timeout to be restarted when new submissions arrive
469 */
470 if (mod_delayed_work(wq: sched->timeout_wq, dwork: &sched->work_tdr, MAX_SCHEDULE_TIMEOUT)
471 && time_after(sched_timeout, now))
472 return sched_timeout - now;
473 else
474 return sched->timeout;
475}
476EXPORT_SYMBOL(drm_sched_suspend_timeout);
477
478/**
479 * drm_sched_resume_timeout - Resume scheduler job timeout
480 *
481 * @sched: scheduler instance for which to resume the timeout
482 * @remaining: remaining timeout
483 *
484 * Resume the delayed work timeout for the scheduler.
485 */
486void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
487 unsigned long remaining)
488{
489 spin_lock(lock: &sched->job_list_lock);
490
491 if (list_empty(head: &sched->pending_list))
492 cancel_delayed_work(dwork: &sched->work_tdr);
493 else
494 mod_delayed_work(wq: sched->timeout_wq, dwork: &sched->work_tdr, delay: remaining);
495
496 spin_unlock(lock: &sched->job_list_lock);
497}
498EXPORT_SYMBOL(drm_sched_resume_timeout);
499
500static void drm_sched_job_begin(struct drm_sched_job *s_job)
501{
502 struct drm_gpu_scheduler *sched = s_job->sched;
503
504 spin_lock(lock: &sched->job_list_lock);
505 list_add_tail(new: &s_job->list, head: &sched->pending_list);
506 drm_sched_start_timeout(sched);
507 spin_unlock(lock: &sched->job_list_lock);
508}
509
510/**
511 * drm_sched_job_reinsert_on_false_timeout - reinsert the job on a false timeout
512 * @sched: scheduler instance
513 * @job: job to be reinserted on the pending list
514 *
515 * In the case of a "false timeout" - when a timeout occurs but the GPU isn't
516 * hung and is making progress, the scheduler must reinsert the job back into
517 * @sched->pending_list. Otherwise, the job and its resources won't be freed
518 * through the &struct drm_sched_backend_ops.free_job callback.
519 *
520 * This function must be used in "false timeout" cases only.
521 */
522static void drm_sched_job_reinsert_on_false_timeout(struct drm_gpu_scheduler *sched,
523 struct drm_sched_job *job)
524{
525 spin_lock(lock: &sched->job_list_lock);
526 list_add(new: &job->list, head: &sched->pending_list);
527
528 /* After reinserting the job, the scheduler enqueues the free-job work
529 * again if ready. Otherwise, a signaled job could be added to the
530 * pending list, but never freed.
531 */
532 drm_sched_run_free_queue(sched);
533 spin_unlock(lock: &sched->job_list_lock);
534}
535
536static void drm_sched_job_timedout(struct work_struct *work)
537{
538 struct drm_gpu_scheduler *sched;
539 struct drm_sched_job *job;
540 enum drm_gpu_sched_stat status = DRM_GPU_SCHED_STAT_RESET;
541
542 sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);
543
544 /* Protects against concurrent deletion in drm_sched_get_finished_job */
545 spin_lock(lock: &sched->job_list_lock);
546 job = list_first_entry_or_null(&sched->pending_list,
547 struct drm_sched_job, list);
548
549 if (job) {
550 /*
551 * Remove the bad job so it cannot be freed by a concurrent
552 * &struct drm_sched_backend_ops.free_job. It will be
553 * reinserted after the scheduler's work items have been
554 * cancelled, at which point it's safe.
555 */
556 list_del_init(entry: &job->list);
557 spin_unlock(lock: &sched->job_list_lock);
558
559 status = job->sched->ops->timedout_job(job);
560
561 /*
562 * Guilty job did complete and hence needs to be manually removed
563 * See drm_sched_stop doc.
564 */
565 if (sched->free_guilty) {
566 job->sched->ops->free_job(job);
567 sched->free_guilty = false;
568 }
569
570 if (status == DRM_GPU_SCHED_STAT_NO_HANG)
571 drm_sched_job_reinsert_on_false_timeout(sched, job);
572 } else {
573 spin_unlock(lock: &sched->job_list_lock);
574 }
575
576 if (status != DRM_GPU_SCHED_STAT_ENODEV)
577 drm_sched_start_timeout_unlocked(sched);
578}
579
580/**
581 * drm_sched_stop - stop the scheduler
582 *
583 * @sched: scheduler instance
584 * @bad: job which caused the time out
585 *
586 * Stop the scheduler and also removes and frees all completed jobs.
587 * Note: bad job will not be freed as it might be used later and so it's
588 * callers responsibility to release it manually if it's not part of the
589 * pending list any more.
590 *
591 * This function is typically used for reset recovery (see the docu of
592 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
593 * scheduler teardown, i.e., before calling drm_sched_fini().
594 *
595 * As it's only used for reset recovery, drivers must not call this function
596 * in their &struct drm_sched_backend_ops.timedout_job callback when they
597 * skip a reset using &enum drm_gpu_sched_stat.DRM_GPU_SCHED_STAT_NO_HANG.
598 */
599void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
600{
601 struct drm_sched_job *s_job, *tmp;
602
603 drm_sched_wqueue_stop(sched);
604
605 /*
606 * Reinsert back the bad job here - now it's safe as
607 * drm_sched_get_finished_job() cannot race against us and release the
608 * bad job at this point - we parked (waited for) any in progress
609 * (earlier) cleanups and drm_sched_get_finished_job() will not be
610 * called now until the scheduler's work items are submitted again.
611 */
612 if (bad && bad->sched == sched)
613 /*
614 * Add at the head of the queue to reflect it was the earliest
615 * job extracted.
616 */
617 list_add(new: &bad->list, head: &sched->pending_list);
618
619 /*
620 * Iterate the job list from later to earlier one and either deactive
621 * their HW callbacks or remove them from pending list if they already
622 * signaled.
623 * This iteration is thread safe as the scheduler's work items have been
624 * cancelled.
625 */
626 list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
627 list) {
628 if (s_job->s_fence->parent &&
629 dma_fence_remove_callback(fence: s_job->s_fence->parent,
630 cb: &s_job->cb)) {
631 dma_fence_put(fence: s_job->s_fence->parent);
632 s_job->s_fence->parent = NULL;
633 atomic_sub(i: s_job->credits, v: &sched->credit_count);
634 } else {
635 /*
636 * remove job from pending_list.
637 * Locking here is for concurrent resume timeout
638 */
639 spin_lock(lock: &sched->job_list_lock);
640 list_del_init(entry: &s_job->list);
641 spin_unlock(lock: &sched->job_list_lock);
642
643 /*
644 * Wait for job's HW fence callback to finish using s_job
645 * before releasing it.
646 *
647 * Job is still alive so fence refcount at least 1
648 */
649 dma_fence_wait(fence: &s_job->s_fence->finished, intr: false);
650
651 /*
652 * We must keep bad job alive for later use during
653 * recovery by some of the drivers but leave a hint
654 * that the guilty job must be released.
655 */
656 if (bad != s_job)
657 sched->ops->free_job(s_job);
658 else
659 sched->free_guilty = true;
660 }
661 }
662
663 /*
664 * Stop pending timer in flight as we rearm it in drm_sched_start. This
665 * avoids the pending timeout work in progress to fire right away after
666 * this TDR finished and before the newly restarted jobs had a
667 * chance to complete.
668 */
669 cancel_delayed_work(dwork: &sched->work_tdr);
670}
671EXPORT_SYMBOL(drm_sched_stop);
672
673/**
674 * drm_sched_start - recover jobs after a reset
675 *
676 * @sched: scheduler instance
677 * @errno: error to set on the pending fences
678 *
679 * This function is typically used for reset recovery (see the docu of
680 * drm_sched_backend_ops.timedout_job() for details). Do not call it for
681 * scheduler startup. The scheduler itself is fully operational after
682 * drm_sched_init() succeeded.
683 *
684 * As it's only used for reset recovery, drivers must not call this function
685 * in their &struct drm_sched_backend_ops.timedout_job callback when they
686 * skip a reset using &enum drm_gpu_sched_stat.DRM_GPU_SCHED_STAT_NO_HANG.
687 */
688void drm_sched_start(struct drm_gpu_scheduler *sched, int errno)
689{
690 struct drm_sched_job *s_job, *tmp;
691
692 /*
693 * Locking the list is not required here as the scheduler's work items
694 * are currently not running, so no new jobs are being inserted or
695 * removed. Also concurrent GPU recovers can't run in parallel.
696 */
697 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
698 struct dma_fence *fence = s_job->s_fence->parent;
699
700 atomic_add(i: s_job->credits, v: &sched->credit_count);
701
702 if (!fence) {
703 drm_sched_job_done(s_job, result: errno ?: -ECANCELED);
704 continue;
705 }
706
707 if (dma_fence_add_callback(fence, cb: &s_job->cb,
708 func: drm_sched_job_done_cb))
709 drm_sched_job_done(s_job, result: fence->error ?: errno);
710 }
711
712 drm_sched_start_timeout_unlocked(sched);
713 drm_sched_wqueue_start(sched);
714}
715EXPORT_SYMBOL(drm_sched_start);
716
717/**
718 * drm_sched_resubmit_jobs - Deprecated, don't use in new code!
719 *
720 * @sched: scheduler instance
721 *
722 * Re-submitting jobs was a concept AMD came up as cheap way to implement
723 * recovery after a job timeout.
724 *
725 * This turned out to be not working very well. First of all there are many
726 * problem with the dma_fence implementation and requirements. Either the
727 * implementation is risking deadlocks with core memory management or violating
728 * documented implementation details of the dma_fence object.
729 *
730 * Drivers can still save and restore their state for recovery operations, but
731 * we shouldn't make this a general scheduler feature around the dma_fence
732 * interface.
733 */
734void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
735{
736 struct drm_sched_job *s_job, *tmp;
737 uint64_t guilty_context;
738 bool found_guilty = false;
739 struct dma_fence *fence;
740
741 list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
742 struct drm_sched_fence *s_fence = s_job->s_fence;
743
744 if (!found_guilty && atomic_read(v: &s_job->karma) > sched->hang_limit) {
745 found_guilty = true;
746 guilty_context = s_job->s_fence->scheduled.context;
747 }
748
749 if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
750 dma_fence_set_error(fence: &s_fence->finished, error: -ECANCELED);
751
752 fence = sched->ops->run_job(s_job);
753
754 if (IS_ERR_OR_NULL(ptr: fence)) {
755 if (IS_ERR(ptr: fence))
756 dma_fence_set_error(fence: &s_fence->finished, error: PTR_ERR(ptr: fence));
757
758 s_job->s_fence->parent = NULL;
759 } else {
760
761 s_job->s_fence->parent = dma_fence_get(fence);
762
763 /* Drop for orignal kref_init */
764 dma_fence_put(fence);
765 }
766 }
767}
768EXPORT_SYMBOL(drm_sched_resubmit_jobs);
769
770/**
771 * drm_sched_job_init - init a scheduler job
772 * @job: scheduler job to init
773 * @entity: scheduler entity to use
774 * @credits: the number of credits this job contributes to the schedulers
775 * credit limit
776 * @owner: job owner for debugging
777 * @drm_client_id: &struct drm_file.client_id of the owner (used by trace
778 * events)
779 *
780 * Refer to drm_sched_entity_push_job() documentation
781 * for locking considerations.
782 *
783 * Drivers must make sure drm_sched_job_cleanup() if this function returns
784 * successfully, even when @job is aborted before drm_sched_job_arm() is called.
785 *
786 * Note that this function does not assign a valid value to each struct member
787 * of struct drm_sched_job. Take a look at that struct's documentation to see
788 * who sets which struct member with what lifetime.
789 *
790 * WARNING: amdgpu abuses &drm_sched.ready to signal when the hardware
791 * has died, which can mean that there's no valid runqueue for a @entity.
792 * This function returns -ENOENT in this case (which probably should be -EIO as
793 * a more meanigful return value).
794 *
795 * Returns 0 for success, negative error code otherwise.
796 */
797int drm_sched_job_init(struct drm_sched_job *job,
798 struct drm_sched_entity *entity,
799 u32 credits, void *owner,
800 uint64_t drm_client_id)
801{
802 if (!entity->rq) {
803 /* This will most likely be followed by missing frames
804 * or worse--a blank screen--leave a trail in the
805 * logs, so this can be debugged easier.
806 */
807 dev_err(job->sched->dev, "%s: entity has no rq!\n", __func__);
808 return -ENOENT;
809 }
810
811 if (unlikely(!credits)) {
812 pr_err("*ERROR* %s: credits cannot be 0!\n", __func__);
813 return -EINVAL;
814 }
815
816 /*
817 * We don't know for sure how the user has allocated. Thus, zero the
818 * struct so that unallowed (i.e., too early) usage of pointers that
819 * this function does not set is guaranteed to lead to a NULL pointer
820 * exception instead of UB.
821 */
822 memset(job, 0, sizeof(*job));
823
824 job->entity = entity;
825 job->credits = credits;
826 job->s_fence = drm_sched_fence_alloc(s_entity: entity, owner, drm_client_id);
827 if (!job->s_fence)
828 return -ENOMEM;
829
830 INIT_LIST_HEAD(list: &job->list);
831
832 xa_init_flags(xa: &job->dependencies, XA_FLAGS_ALLOC);
833
834 return 0;
835}
836EXPORT_SYMBOL(drm_sched_job_init);
837
838/**
839 * drm_sched_job_arm - arm a scheduler job for execution
840 * @job: scheduler job to arm
841 *
842 * This arms a scheduler job for execution. Specifically it initializes the
843 * &drm_sched_job.s_fence of @job, so that it can be attached to struct dma_resv
844 * or other places that need to track the completion of this job. It also
845 * initializes sequence numbers, which are fundamental for fence ordering.
846 *
847 * Refer to drm_sched_entity_push_job() documentation for locking
848 * considerations.
849 *
850 * Once this function was called, you *must* submit @job with
851 * drm_sched_entity_push_job().
852 *
853 * This can only be called if drm_sched_job_init() succeeded.
854 */
855void drm_sched_job_arm(struct drm_sched_job *job)
856{
857 struct drm_gpu_scheduler *sched;
858 struct drm_sched_entity *entity = job->entity;
859
860 BUG_ON(!entity);
861 drm_sched_entity_select_rq(entity);
862 sched = entity->rq->sched;
863
864 job->sched = sched;
865 job->s_priority = entity->priority;
866
867 drm_sched_fence_init(fence: job->s_fence, entity: job->entity);
868}
869EXPORT_SYMBOL(drm_sched_job_arm);
870
871/**
872 * drm_sched_job_add_dependency - adds the fence as a job dependency
873 * @job: scheduler job to add the dependencies to
874 * @fence: the dma_fence to add to the list of dependencies.
875 *
876 * Note that @fence is consumed in both the success and error cases.
877 *
878 * Returns:
879 * 0 on success, or an error on failing to expand the array.
880 */
881int drm_sched_job_add_dependency(struct drm_sched_job *job,
882 struct dma_fence *fence)
883{
884 struct dma_fence *entry;
885 unsigned long index;
886 u32 id = 0;
887 int ret;
888
889 if (!fence)
890 return 0;
891
892 /* Deduplicate if we already depend on a fence from the same context.
893 * This lets the size of the array of deps scale with the number of
894 * engines involved, rather than the number of BOs.
895 */
896 xa_for_each(&job->dependencies, index, entry) {
897 if (entry->context != fence->context)
898 continue;
899
900 if (dma_fence_is_later(f1: fence, f2: entry)) {
901 dma_fence_put(fence: entry);
902 xa_store(&job->dependencies, index, entry: fence, GFP_KERNEL);
903 } else {
904 dma_fence_put(fence);
905 }
906 return 0;
907 }
908
909 ret = xa_alloc(xa: &job->dependencies, id: &id, entry: fence, xa_limit_32b, GFP_KERNEL);
910 if (ret != 0)
911 dma_fence_put(fence);
912
913 return ret;
914}
915EXPORT_SYMBOL(drm_sched_job_add_dependency);
916
917/**
918 * drm_sched_job_add_syncobj_dependency - adds a syncobj's fence as a job dependency
919 * @job: scheduler job to add the dependencies to
920 * @file: drm file private pointer
921 * @handle: syncobj handle to lookup
922 * @point: timeline point
923 *
924 * This adds the fence matching the given syncobj to @job.
925 *
926 * Returns:
927 * 0 on success, or an error on failing to expand the array.
928 */
929int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job,
930 struct drm_file *file,
931 u32 handle,
932 u32 point)
933{
934 struct dma_fence *fence;
935 int ret;
936
937 ret = drm_syncobj_find_fence(file_private: file, handle, point, flags: 0, fence: &fence);
938 if (ret)
939 return ret;
940
941 return drm_sched_job_add_dependency(job, fence);
942}
943EXPORT_SYMBOL(drm_sched_job_add_syncobj_dependency);
944
945/**
946 * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
947 * @job: scheduler job to add the dependencies to
948 * @resv: the dma_resv object to get the fences from
949 * @usage: the dma_resv_usage to use to filter the fences
950 *
951 * This adds all fences matching the given usage from @resv to @job.
952 * Must be called with the @resv lock held.
953 *
954 * Returns:
955 * 0 on success, or an error on failing to expand the array.
956 */
957int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job,
958 struct dma_resv *resv,
959 enum dma_resv_usage usage)
960{
961 struct dma_resv_iter cursor;
962 struct dma_fence *fence;
963 int ret;
964
965 dma_resv_assert_held(resv);
966
967 dma_resv_for_each_fence(&cursor, resv, usage, fence) {
968 /*
969 * As drm_sched_job_add_dependency always consumes the fence
970 * reference (even when it fails), and dma_resv_for_each_fence
971 * is not obtaining one, we need to grab one before calling.
972 */
973 ret = drm_sched_job_add_dependency(job, dma_fence_get(fence));
974 if (ret)
975 return ret;
976 }
977 return 0;
978}
979EXPORT_SYMBOL(drm_sched_job_add_resv_dependencies);
980
981/**
982 * drm_sched_job_add_implicit_dependencies - adds implicit dependencies as job
983 * dependencies
984 * @job: scheduler job to add the dependencies to
985 * @obj: the gem object to add new dependencies from.
986 * @write: whether the job might write the object (so we need to depend on
987 * shared fences in the reservation object).
988 *
989 * This should be called after drm_gem_lock_reservations() on your array of
990 * GEM objects used in the job but before updating the reservations with your
991 * own fences.
992 *
993 * Returns:
994 * 0 on success, or an error on failing to expand the array.
995 */
996int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
997 struct drm_gem_object *obj,
998 bool write)
999{
1000 return drm_sched_job_add_resv_dependencies(job, obj->resv,
1001 dma_resv_usage_rw(write));
1002}
1003EXPORT_SYMBOL(drm_sched_job_add_implicit_dependencies);
1004
1005/**
1006 * drm_sched_job_has_dependency - check whether fence is the job's dependency
1007 * @job: scheduler job to check
1008 * @fence: fence to look for
1009 *
1010 * Returns:
1011 * True if @fence is found within the job's dependencies, or otherwise false.
1012 */
1013bool drm_sched_job_has_dependency(struct drm_sched_job *job,
1014 struct dma_fence *fence)
1015{
1016 struct dma_fence *f;
1017 unsigned long index;
1018
1019 xa_for_each(&job->dependencies, index, f) {
1020 if (f == fence)
1021 return true;
1022 }
1023
1024 return false;
1025}
1026EXPORT_SYMBOL(drm_sched_job_has_dependency);
1027
1028/**
1029 * drm_sched_job_cleanup - clean up scheduler job resources
1030 * @job: scheduler job to clean up
1031 *
1032 * Cleans up the resources allocated with drm_sched_job_init().
1033 *
1034 * Drivers should call this from their error unwind code if @job is aborted
1035 * before drm_sched_job_arm() is called.
1036 *
1037 * drm_sched_job_arm() is a point of no return since it initializes the fences
1038 * and their sequence number etc. Once that function has been called, you *must*
1039 * submit it with drm_sched_entity_push_job() and cannot simply abort it by
1040 * calling drm_sched_job_cleanup().
1041 *
1042 * This function should be called in the &drm_sched_backend_ops.free_job callback.
1043 */
1044void drm_sched_job_cleanup(struct drm_sched_job *job)
1045{
1046 struct dma_fence *fence;
1047 unsigned long index;
1048
1049 if (kref_read(kref: &job->s_fence->finished.refcount)) {
1050 /* The job has been processed by the scheduler, i.e.,
1051 * drm_sched_job_arm() and drm_sched_entity_push_job() have
1052 * been called.
1053 */
1054 dma_fence_put(fence: &job->s_fence->finished);
1055 } else {
1056 /* The job was aborted before it has been committed to be run;
1057 * notably, drm_sched_job_arm() has not been called.
1058 */
1059 drm_sched_fence_free(fence: job->s_fence);
1060 }
1061
1062 job->s_fence = NULL;
1063
1064 xa_for_each(&job->dependencies, index, fence) {
1065 dma_fence_put(fence);
1066 }
1067 xa_destroy(&job->dependencies);
1068
1069}
1070EXPORT_SYMBOL(drm_sched_job_cleanup);
1071
1072/**
1073 * drm_sched_wakeup - Wake up the scheduler if it is ready to queue
1074 * @sched: scheduler instance
1075 *
1076 * Wake up the scheduler if we can queue jobs.
1077 */
1078void drm_sched_wakeup(struct drm_gpu_scheduler *sched)
1079{
1080 drm_sched_run_job_queue(sched);
1081}
1082
1083/**
1084 * drm_sched_select_entity - Select next entity to process
1085 *
1086 * @sched: scheduler instance
1087 *
1088 * Return an entity to process or NULL if none are found.
1089 *
1090 * Note, that we break out of the for-loop when "entity" is non-null, which can
1091 * also be an error-pointer--this assures we don't process lower priority
1092 * run-queues. See comments in the respectively called functions.
1093 */
1094static struct drm_sched_entity *
1095drm_sched_select_entity(struct drm_gpu_scheduler *sched)
1096{
1097 struct drm_sched_entity *entity;
1098 int i;
1099
1100 /* Start with the highest priority.
1101 */
1102 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1103 entity = drm_sched_policy == DRM_SCHED_POLICY_FIFO ?
1104 drm_sched_rq_select_entity_fifo(sched, rq: sched->sched_rq[i]) :
1105 drm_sched_rq_select_entity_rr(sched, rq: sched->sched_rq[i]);
1106 if (entity)
1107 break;
1108 }
1109
1110 return IS_ERR(ptr: entity) ? NULL : entity;
1111}
1112
1113/**
1114 * drm_sched_get_finished_job - fetch the next finished job to be destroyed
1115 *
1116 * @sched: scheduler instance
1117 * @have_more: are there more finished jobs on the list
1118 *
1119 * Informs the caller through @have_more whether there are more finished jobs
1120 * besides the returned one.
1121 *
1122 * Returns the next finished job from the pending list (if there is one)
1123 * ready for it to be destroyed.
1124 */
1125static struct drm_sched_job *
1126drm_sched_get_finished_job(struct drm_gpu_scheduler *sched, bool *have_more)
1127{
1128 struct drm_sched_job *job, *next;
1129
1130 spin_lock(lock: &sched->job_list_lock);
1131
1132 job = list_first_entry_or_null(&sched->pending_list,
1133 struct drm_sched_job, list);
1134 if (job && dma_fence_is_signaled(fence: &job->s_fence->finished)) {
1135 /* remove job from pending_list */
1136 list_del_init(entry: &job->list);
1137
1138 /* cancel this job's TO timer */
1139 cancel_delayed_work(dwork: &sched->work_tdr);
1140
1141 *have_more = false;
1142 next = list_first_entry_or_null(&sched->pending_list,
1143 typeof(*next), list);
1144 if (next) {
1145 /* make the scheduled timestamp more accurate */
1146 if (test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT,
1147 &next->s_fence->scheduled.flags))
1148 next->s_fence->scheduled.timestamp =
1149 dma_fence_timestamp(fence: &job->s_fence->finished);
1150
1151 *have_more = dma_fence_is_signaled(fence: &next->s_fence->finished);
1152
1153 /* start TO timer for next job */
1154 drm_sched_start_timeout(sched);
1155 }
1156 } else {
1157 job = NULL;
1158 }
1159
1160 spin_unlock(lock: &sched->job_list_lock);
1161
1162 return job;
1163}
1164
1165/**
1166 * drm_sched_pick_best - Get a drm sched from a sched_list with the least load
1167 * @sched_list: list of drm_gpu_schedulers
1168 * @num_sched_list: number of drm_gpu_schedulers in the sched_list
1169 *
1170 * Returns pointer of the sched with the least load or NULL if none of the
1171 * drm_gpu_schedulers are ready
1172 */
1173struct drm_gpu_scheduler *
1174drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
1175 unsigned int num_sched_list)
1176{
1177 struct drm_gpu_scheduler *sched, *picked_sched = NULL;
1178 int i;
1179 unsigned int min_score = UINT_MAX, num_score;
1180
1181 for (i = 0; i < num_sched_list; ++i) {
1182 sched = sched_list[i];
1183
1184 if (!sched->ready) {
1185 DRM_WARN("scheduler %s is not ready, skipping",
1186 sched->name);
1187 continue;
1188 }
1189
1190 num_score = atomic_read(v: sched->score);
1191 if (num_score < min_score) {
1192 min_score = num_score;
1193 picked_sched = sched;
1194 }
1195 }
1196
1197 return picked_sched;
1198}
1199EXPORT_SYMBOL(drm_sched_pick_best);
1200
1201/**
1202 * drm_sched_free_job_work - worker to call free_job
1203 *
1204 * @w: free job work
1205 */
1206static void drm_sched_free_job_work(struct work_struct *w)
1207{
1208 struct drm_gpu_scheduler *sched =
1209 container_of(w, struct drm_gpu_scheduler, work_free_job);
1210 struct drm_sched_job *job;
1211 bool have_more;
1212
1213 job = drm_sched_get_finished_job(sched, have_more: &have_more);
1214 if (job) {
1215 sched->ops->free_job(job);
1216 if (have_more)
1217 drm_sched_run_free_queue(sched);
1218 }
1219
1220 drm_sched_run_job_queue(sched);
1221}
1222
1223/**
1224 * drm_sched_run_job_work - worker to call run_job
1225 *
1226 * @w: run job work
1227 */
1228static void drm_sched_run_job_work(struct work_struct *w)
1229{
1230 struct drm_gpu_scheduler *sched =
1231 container_of(w, struct drm_gpu_scheduler, work_run_job);
1232 struct drm_sched_entity *entity;
1233 struct dma_fence *fence;
1234 struct drm_sched_fence *s_fence;
1235 struct drm_sched_job *sched_job;
1236 int r;
1237
1238 /* Find entity with a ready job */
1239 entity = drm_sched_select_entity(sched);
1240 if (!entity) {
1241 /*
1242 * Either no more work to do, or the next ready job needs more
1243 * credits than the scheduler has currently available.
1244 */
1245 return;
1246 }
1247
1248 sched_job = drm_sched_entity_pop_job(entity);
1249 if (!sched_job) {
1250 complete_all(&entity->entity_idle);
1251 drm_sched_run_job_queue(sched);
1252 return;
1253 }
1254
1255 s_fence = sched_job->s_fence;
1256
1257 atomic_add(i: sched_job->credits, v: &sched->credit_count);
1258 drm_sched_job_begin(s_job: sched_job);
1259
1260 trace_drm_sched_job_run(sched_job, entity);
1261 /*
1262 * The run_job() callback must by definition return a fence whose
1263 * refcount has been incremented for the scheduler already.
1264 */
1265 fence = sched->ops->run_job(sched_job);
1266 complete_all(&entity->entity_idle);
1267 drm_sched_fence_scheduled(fence: s_fence, parent: fence);
1268
1269 if (!IS_ERR_OR_NULL(ptr: fence)) {
1270 r = dma_fence_add_callback(fence, cb: &sched_job->cb,
1271 func: drm_sched_job_done_cb);
1272 if (r == -ENOENT)
1273 drm_sched_job_done(s_job: sched_job, result: fence->error);
1274 else if (r)
1275 DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
1276
1277 dma_fence_put(fence);
1278 } else {
1279 drm_sched_job_done(s_job: sched_job, result: IS_ERR(ptr: fence) ?
1280 PTR_ERR(ptr: fence) : 0);
1281 }
1282
1283 wake_up(&sched->job_scheduled);
1284 drm_sched_run_job_queue(sched);
1285}
1286
1287static struct workqueue_struct *drm_sched_alloc_wq(const char *name)
1288{
1289#if (IS_ENABLED(CONFIG_LOCKDEP))
1290 static struct lockdep_map map = {
1291 .name = "drm_sched_lockdep_map"
1292 };
1293
1294 /*
1295 * Avoid leaking a lockdep map on each drm sched creation and
1296 * destruction by using a single lockdep map for all drm sched
1297 * allocated submit_wq.
1298 */
1299
1300 return alloc_ordered_workqueue_lockdep_map(name, WQ_MEM_RECLAIM, &map);
1301#else
1302 return alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
1303#endif
1304}
1305
1306/**
1307 * drm_sched_init - Init a gpu scheduler instance
1308 *
1309 * @sched: scheduler instance
1310 * @args: scheduler initialization arguments
1311 *
1312 * Return 0 on success, otherwise error code.
1313 */
1314int drm_sched_init(struct drm_gpu_scheduler *sched, const struct drm_sched_init_args *args)
1315{
1316 int i;
1317
1318 sched->ops = args->ops;
1319 sched->credit_limit = args->credit_limit;
1320 sched->name = args->name;
1321 sched->timeout = args->timeout;
1322 sched->hang_limit = args->hang_limit;
1323 sched->timeout_wq = args->timeout_wq ? args->timeout_wq : system_percpu_wq;
1324 sched->score = args->score ? args->score : &sched->_score;
1325 sched->dev = args->dev;
1326
1327 if (args->num_rqs > DRM_SCHED_PRIORITY_COUNT) {
1328 /* This is a gross violation--tell drivers what the problem is.
1329 */
1330 dev_err(sched->dev, "%s: num_rqs cannot be greater than DRM_SCHED_PRIORITY_COUNT\n",
1331 __func__);
1332 return -EINVAL;
1333 } else if (sched->sched_rq) {
1334 /* Not an error, but warn anyway so drivers can
1335 * fine-tune their DRM calling order, and return all
1336 * is good.
1337 */
1338 dev_warn(sched->dev, "%s: scheduler already initialized!\n", __func__);
1339 return 0;
1340 }
1341
1342 if (args->submit_wq) {
1343 sched->submit_wq = args->submit_wq;
1344 sched->own_submit_wq = false;
1345 } else {
1346 sched->submit_wq = drm_sched_alloc_wq(name: args->name);
1347 if (!sched->submit_wq)
1348 return -ENOMEM;
1349
1350 sched->own_submit_wq = true;
1351 }
1352
1353 sched->sched_rq = kmalloc_array(args->num_rqs, sizeof(*sched->sched_rq),
1354 GFP_KERNEL | __GFP_ZERO);
1355 if (!sched->sched_rq)
1356 goto Out_check_own;
1357 sched->num_rqs = args->num_rqs;
1358 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1359 sched->sched_rq[i] = kzalloc(sizeof(*sched->sched_rq[i]), GFP_KERNEL);
1360 if (!sched->sched_rq[i])
1361 goto Out_unroll;
1362 drm_sched_rq_init(sched, rq: sched->sched_rq[i]);
1363 }
1364
1365 init_waitqueue_head(&sched->job_scheduled);
1366 INIT_LIST_HEAD(list: &sched->pending_list);
1367 spin_lock_init(&sched->job_list_lock);
1368 atomic_set(v: &sched->credit_count, i: 0);
1369 INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
1370 INIT_WORK(&sched->work_run_job, drm_sched_run_job_work);
1371 INIT_WORK(&sched->work_free_job, drm_sched_free_job_work);
1372 atomic_set(v: &sched->_score, i: 0);
1373 atomic64_set(v: &sched->job_id_count, i: 0);
1374 sched->pause_submit = false;
1375
1376 sched->ready = true;
1377 return 0;
1378Out_unroll:
1379 for (--i ; i >= DRM_SCHED_PRIORITY_KERNEL; i--)
1380 kfree(objp: sched->sched_rq[i]);
1381
1382 kfree(objp: sched->sched_rq);
1383 sched->sched_rq = NULL;
1384Out_check_own:
1385 if (sched->own_submit_wq)
1386 destroy_workqueue(wq: sched->submit_wq);
1387 dev_err(sched->dev, "%s: Failed to setup GPU scheduler--out of memory\n", __func__);
1388 return -ENOMEM;
1389}
1390EXPORT_SYMBOL(drm_sched_init);
1391
1392static void drm_sched_cancel_remaining_jobs(struct drm_gpu_scheduler *sched)
1393{
1394 struct drm_sched_job *job, *tmp;
1395
1396 /* All other accessors are stopped. No locking necessary. */
1397 list_for_each_entry_safe_reverse(job, tmp, &sched->pending_list, list) {
1398 sched->ops->cancel_job(job);
1399 list_del(entry: &job->list);
1400 sched->ops->free_job(job);
1401 }
1402}
1403
1404/**
1405 * drm_sched_fini - Destroy a gpu scheduler
1406 *
1407 * @sched: scheduler instance
1408 *
1409 * Tears down and cleans up the scheduler.
1410 *
1411 * This stops submission of new jobs to the hardware through &struct
1412 * drm_sched_backend_ops.run_job. If &struct drm_sched_backend_ops.cancel_job
1413 * is implemented, all jobs will be canceled through it and afterwards cleaned
1414 * up through &struct drm_sched_backend_ops.free_job. If cancel_job is not
1415 * implemented, memory could leak.
1416 */
1417void drm_sched_fini(struct drm_gpu_scheduler *sched)
1418{
1419 struct drm_sched_entity *s_entity;
1420 int i;
1421
1422 drm_sched_wqueue_stop(sched);
1423
1424 for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
1425 struct drm_sched_rq *rq = sched->sched_rq[i];
1426
1427 spin_lock(lock: &rq->lock);
1428 list_for_each_entry(s_entity, &rq->entities, list) {
1429 /*
1430 * Prevents reinsertion and marks job_queue as idle,
1431 * it will be removed from the rq in drm_sched_entity_fini()
1432 * eventually
1433 *
1434 * FIXME:
1435 * This lacks the proper spin_lock(&s_entity->lock) and
1436 * is, therefore, a race condition. Most notably, it
1437 * can race with drm_sched_entity_push_job(). The lock
1438 * cannot be taken here, however, because this would
1439 * lead to lock inversion -> deadlock.
1440 *
1441 * The best solution probably is to enforce the life
1442 * time rule of all entities having to be torn down
1443 * before their scheduler. Then, however, locking could
1444 * be dropped alltogether from this function.
1445 *
1446 * For now, this remains a potential race in all
1447 * drivers that keep entities alive for longer than
1448 * the scheduler.
1449 *
1450 * The READ_ONCE() is there to make the lockless read
1451 * (warning about the lockless write below) slightly
1452 * less broken...
1453 */
1454 if (!READ_ONCE(s_entity->stopped))
1455 dev_warn(sched->dev, "Tearing down scheduler with active entities!\n");
1456 s_entity->stopped = true;
1457 }
1458 spin_unlock(lock: &rq->lock);
1459 kfree(objp: sched->sched_rq[i]);
1460 }
1461
1462 /* Wakeup everyone stuck in drm_sched_entity_flush for this scheduler */
1463 wake_up_all(&sched->job_scheduled);
1464
1465 /* Confirm no work left behind accessing device structures */
1466 cancel_delayed_work_sync(dwork: &sched->work_tdr);
1467
1468 /* Avoid memory leaks if supported by the driver. */
1469 if (sched->ops->cancel_job)
1470 drm_sched_cancel_remaining_jobs(sched);
1471
1472 if (sched->own_submit_wq)
1473 destroy_workqueue(wq: sched->submit_wq);
1474 sched->ready = false;
1475 kfree(objp: sched->sched_rq);
1476 sched->sched_rq = NULL;
1477
1478 if (!list_empty(head: &sched->pending_list))
1479 dev_warn(sched->dev, "Tearing down scheduler while jobs are pending!\n");
1480}
1481EXPORT_SYMBOL(drm_sched_fini);
1482
1483/**
1484 * drm_sched_increase_karma - Update sched_entity guilty flag
1485 *
1486 * @bad: The job guilty of time out
1487 *
1488 * Increment on every hang caused by the 'bad' job. If this exceeds the hang
1489 * limit of the scheduler then the respective sched entity is marked guilty and
1490 * jobs from it will not be scheduled further
1491 */
1492void drm_sched_increase_karma(struct drm_sched_job *bad)
1493{
1494 int i;
1495 struct drm_sched_entity *tmp;
1496 struct drm_sched_entity *entity;
1497 struct drm_gpu_scheduler *sched = bad->sched;
1498
1499 /* don't change @bad's karma if it's from KERNEL RQ,
1500 * because sometimes GPU hang would cause kernel jobs (like VM updating jobs)
1501 * corrupt but keep in mind that kernel jobs always considered good.
1502 */
1503 if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
1504 atomic_inc(v: &bad->karma);
1505
1506 for (i = DRM_SCHED_PRIORITY_HIGH; i < sched->num_rqs; i++) {
1507 struct drm_sched_rq *rq = sched->sched_rq[i];
1508
1509 spin_lock(lock: &rq->lock);
1510 list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
1511 if (bad->s_fence->scheduled.context ==
1512 entity->fence_context) {
1513 if (entity->guilty)
1514 atomic_set(v: entity->guilty, i: 1);
1515 break;
1516 }
1517 }
1518 spin_unlock(lock: &rq->lock);
1519 if (&entity->list != &rq->entities)
1520 break;
1521 }
1522 }
1523}
1524EXPORT_SYMBOL(drm_sched_increase_karma);
1525
1526/**
1527 * drm_sched_wqueue_ready - Is the scheduler ready for submission
1528 *
1529 * @sched: scheduler instance
1530 *
1531 * Returns true if submission is ready
1532 */
1533bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched)
1534{
1535 return sched->ready;
1536}
1537EXPORT_SYMBOL(drm_sched_wqueue_ready);
1538
1539/**
1540 * drm_sched_wqueue_stop - stop scheduler submission
1541 * @sched: scheduler instance
1542 *
1543 * Stops the scheduler from pulling new jobs from entities. It also stops
1544 * freeing jobs automatically through drm_sched_backend_ops.free_job().
1545 */
1546void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched)
1547{
1548 WRITE_ONCE(sched->pause_submit, true);
1549 cancel_work_sync(work: &sched->work_run_job);
1550 cancel_work_sync(work: &sched->work_free_job);
1551}
1552EXPORT_SYMBOL(drm_sched_wqueue_stop);
1553
1554/**
1555 * drm_sched_wqueue_start - start scheduler submission
1556 * @sched: scheduler instance
1557 *
1558 * Restarts the scheduler after drm_sched_wqueue_stop() has stopped it.
1559 *
1560 * This function is not necessary for 'conventional' startup. The scheduler is
1561 * fully operational after drm_sched_init() succeeded.
1562 */
1563void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched)
1564{
1565 WRITE_ONCE(sched->pause_submit, false);
1566 queue_work(wq: sched->submit_wq, work: &sched->work_run_job);
1567 queue_work(wq: sched->submit_wq, work: &sched->work_free_job);
1568}
1569EXPORT_SYMBOL(drm_sched_wqueue_start);
1570

source code of linux/drivers/gpu/drm/scheduler/sched_main.c